一、软件介绍
文末提供程序和源码下载
众所周知,用于监督标记问题的深度神经网络 (DNN) 可以在各种学习任务中产生准确的结果。但是,当准确性是唯一目标时,DNN 经常会做出过于自信的预测,并且无论测试数据是否属于任何已知标签,它们也总是进行标签预测。
EQUINE was created to simplify two kinds of uncertainty quantification for supervised labeling problems:
EQUINE 的创建是为了简化监督标记问题的两种不确定性量化:
- Calibrated probabilities for each predicted label
每个预测标签的校准概率 - An in-distribution score, indicating whether any of the model's known labels should be trusted.
分布内分数,指示是否应信任模型的任何已知标签。
二、Installation 安装
Users are recommended to install a virtual environment such as Anaconda, as is also recommended in the pytorch installation. EQUINE has relatively few dependencies beyond torch.
建议用户安装虚拟环境,例如 Anaconda,pytorch 安装中也建议安装。EQUINE 除了 torch 之外的依赖项相对较少。
pip install equine
Design 设计
EQUINE extends pytorch's nn.Module
interface using a predict
method that returns both the class predictions and the extra OOD scores.
EQUINE 使用一种 predict
返回类预测和额外 OOD 分数的方法扩展了 nn.Module
pytorch 的接口。
三、软件下载
迅雷云盘
本文信息来源于GitHub作者地址:https://github.com/mit-ll-responsible-ai/equine